The Corpus: What is the difference between The Party Plalist and The Mellow Playlist?

The Playlists

The Playlists


As a teenager, I used to be the “music guy” of my group. I felt like I had the responsibility of to ensure good music during our gatherings. This seems like an easy job but it can get pretty hard. Besides putting hours into finding good music, I had to play the right song at the right time to fit the energy. This is often overlooked but timing is crucial, as the atmosphere of the group can make or break the song, even if it is an amazing track in itself. I realised however, that always looking for the next song can get tiring. As a result, I made two playlists, a Party and Mellow playlist, for two different occasions. This way I did not have to worry about what song was going play next.

For my corpus, I decided to analyse these two playlist. The Party Playlist mainly contains commercial hiphop but occasionally some famous pop songs by artist such as Bruno Mars and Beyonce. The Mellow Playlist however, is more of an amalgam of genres as it contains Jazz, Hiphop, Neo-soul and Indie. As previously mentioned, I created these playlists in order to serve two very different occasions. Consequently, the sonic differences in these playlist are clearly audible. In the following storyboards I will analyse the sonic differences between the two playlists through the Spotify API, and determine Spotify perceives the same sonic differences as I do.

Comparing The Danceability of The Playlists


The goal of The Party Playlist is get people dancing. For The Mellow Playlist, it is to fit a calm but positive atmosphere. This mayor difference led me to the first analyses on one specific feature: danceability.

Firstly, I wanted see whether the difference in danceability was influenced by different features of the tracks. With a tempo an average tempo 122BPM, the party playlist surpasses the 105 BPM of the mellow playlist. My initial guess was that tempo was a heavily influenced the danceability value. Looking at the plot “Analysing The Features That Influence Danceability” this relationship is visible. The danceability decreases quickly after the BPM goes below the 100. I added the energy feature, in an attempt to find a pattern between the three features. There seems to be a slightly stronger relationship between the energy and danceability but even this one is not robust.

Secondly, I created another graph in order to compare the mean danceability of the playlists. This is displayed by the black badge on the graph. The difference between playlist’s danceability is minor, and it becomes even smaller when adding or subtracting the standard deviation. This surprised me as the sonic differences are huge in my opinion.

Playlist Mean Danceability Standard Deviation
The Party Playlist 0.789 0.110
The Mellow Playlist 0.607 0.146

Comparing In Each Playlist The Track With the Mean Danceability To The Outlier


In an attempt to delve deeper into the danceability feature, I have selected from each playlist a track with a danceability value close to the mean of its respective playlist. Secondly, from each playlist I took a track with a danceability value which deviates strongly from the mean. This means that from The Mellow Playlist, I have chosen a track with the highest danceability and from The Party Playlist a track with a lowes danceability. I will analyse and compare these track to the average track of the playlist. By doing this, I hope to discover why, despite being placed (by me) in the playlist, the songs have such a deviating danceability. The mentioned tracks have been put in a table with its name, respective playlist and danceability value:

Playlist Name Danceability
The Party Playlist 4/5 0.789
The Party Playlist Hoofdprijs 0.424
The Mellow Playlist Amphetamine 0.610
The Mellow Playlist Lost 0.913

Since Lost and 4/5, and Amphetamine and Hoofdprijs have a close danceability, I’m expecting the cepstograms to show some similarity despite coming from different playlists.

Analysis Of Cepstograms


The cepstograms beautifully display the similarity of Lost and 4/5, and Amphetamine and Hoofdprijs. Despite coming from the same playlist, Lost and Amphetamine differ greatly from each other when looking at the overall intensity of the timbre features. There are a few interesting timbre features displayed in the both cepstograms. When looking at Lost’s, the brightest regions are in the first 35, 90 and 220 seconds at feature c04. According to Spotify c04 represents sounds with a high attack. This would make sense as these peaks coincide exactly with the calmer points in the song consisting of drums, light guitar and sometimes vocals. Amphetamine’s cepstogram is less bright than Lost’s. Still, it contain some interesting points, for example at 175 seconds in c02 and c03. According to Spotify, c02 and c03 indicate brightness and flatness of a sound, respectively. Around that time, the sound is transitioning into the second part of the track, which is more hiphop. Musically, this is done with a high pitched, low fidelity sounding organ while the previous song sharply fades out. The spike in c02 can be attributed to this fade out. Lastly, the brightnes of c03 around the same time can be attributed to the complete transition, which takes around 20 seconds.

The cepstogram of 4/5 is the brightest. This surprises me as musically, it is not that interesting. It has 2 two other sounds besides the drums. Furthermore, the pattern of these sounds are simple and repeated throughout the song. Furthermore, the vocals of 4/5 contain autotune and could definitely influence the ooveral timbre of the sound Lastly, this could be due to the loudness of the track, the cepstogram is registering high energy on every timbre. feature.

The last track is Hoofprijs by Yung Nnelg. This is in my opinion a pretty high energy track however, the cepstogram is not registering it that way. Comparing this track to 4/5 is that the main sound of this track, as opposed to the high pitched and bright sounding riff in 4/5, is low pitch wise and contains a dark sounding piano. Furthermore, the vocals as opposed to 4/5 is dark, low and does not contain any autotune.

These results suggest that, as seen with Lost and 4/5, a high energy distribution across timbres influence the danceability.

Timbre SSM of The Four Songs


I was hoping to see some some similarities between Lost and 4/5 as they both have a high danceability. This is however not the case. Looking at the Self-Similarity matrices, they are all very different. When looking at Hoofdprijs, there seems to be a vertical (and horizontal) line at around 180 seconds. Similarly Amphetamine contains one as well. When listening to these two tracks, the bright yellow lines correspond to a dip in musical intensity. For Hoofdprijs, this is an interlude with a strong synth but no drums. For Amphetamine this is the transition discussed in the Analysis of Cepstograms. This dip in musical intensity could explain why Hoofdprijs has a danceability comparable to that of Amphetamine, instead of 4/5. I was however, expecting a different patteren after the musical transition of Amphetamine. When looking closer at the Self-Similarity Matrix of lost, there are three fat yellow lines that divide the matrix. These parts, are the chorus of the song which occurs at the beginning, around the middle and at the end. The bluer parts are the verses in Lost. Looking at 4/5’s Self- Similarity Matrix, there is an obvious change in timbre at around 80 seconds. This can be attributed to the addition of an instrument and a bridge. As opposed the matrices of Hoofprijs and Amphetamine, 4/5’s and Lost’s matrices are not comparable. Overall, I observed no patterns that suggest a link danceability.

Analysis of tempograms of the four songs


When creating my playlists, I would assign a track to the right playlist based on its energy. Now, energy is a subjective term and spotify has its own definition for it. In my definition as well as in spotify’s, tempo plays a large factor. The tempograms allow me to visualise how the tempo of a track is felt. The emphasis is on ‘felt’ as subconsciously humans perceive, besides the normal tempo, the tempo pulses. The listener hears triplets, quarter notes, tatums etc. These all add to the feeling of the tempo.

Looking at the generated tempograms, 4/5 and Hoofdprijs seem the most consisted. As they are both hiphop, they have a consistent and danceable tempo, as opposed to that of Lost and Amphetamine. This is a reason for why I choose to put 4/5 and Hoofdprijs into The Party Playlist. Another interesting point, is the perceived tempo of 4/5 and Lost. The line with the highest energy, is in both graphs at around 310 BPM. Without doubt, the value is too high for it to be the actual tempo, so it must be a tempo harmonic. This similarity is less noticeable when looking at the tempograms of Amphetamine and Hoofdprijs. With Amphetamine, the lower line at the start of the song is at a comparable tempo to that of Hoofdprijs’s. Amphetamine however, has another line at around 350 BPM which could be an explanation for why Amphetamine has a higher danceability than Hoofdprijs.

Analysis of the Chordograms the four songs


Even though this corpus is mainly focused on danceability, it is interesting to look at the chordograms. My assumption is that chord changes result a higher energy, which in turn can affect the danceability.

Lost has a lot of energy distributed over a large amount of chords. The most prominent one is Bb7, with as second strongest Amin. Musically, this does not make a lot of sense as the roots are a half step apart. When looking at Hoofdprijs, the energy distribution across the chords is more clear. Here as well, the most prominent chord is Bb7. However, E major and E7 appear together as well, which just like with Lost is musically not probable. Amphetamine and 4/5 seem to be experiencing the same problem with clashing chords. Unfortunately this means that it is impossible to draw a good conclusion from these graphs.

Training a classifier.

Classifier Prediction

Classifier Prediction


Since I made the playlists based on how I perceived different tracks, it would be interesting to see how spotify would divide the tracks across the two playlists. In order to do that, I made a confusion matrix. I amended the feature selection by removing valence and duration, as my corpus mainly focused on danceability, tempo, timbre and energy. Lastly, in order to increase the accuracy, I also increased to fold validation to 10.

Prediction The Mellow Playlist The Party Playlist
The Mellow Playlist 119 31
The Party Playlist 22 110

I used the graph in order to illustrate the difference in timbre between the two playlist. A notable difference is that The Party Playlist is more to the right, meaning that it is louder. This makes sense as party songs are usually made for clubs and have a stronger bass.


Conclusion

Analysing my two playlists has given me some interesting insights about how the Spotify API analyses music. My goal was see whether the sonic differences which influenced my track choice for my playlist, were also perceived by Spotify. Since the biggest factor for my track choice for my playlist is danceability, I narrowed down on that feature. The results that I got were not as clear as I had hoped. Nevertheless, a few conclusions can be drawn. Firstly, tempo is a determining factor for danceability. The first graph allowed me to see this relationship. There is a visible decrease of danceability after the tempo decreases below 100 BPM. Secondly, tempo overtones also influence the danceability. This conclusion was drawn when looking at the tempograms. Even though I placed Amphetamine into The Mellow Playlist, it had a higher danceability than Hoofdprijs. I concluded that it was a result of the higher tempo overtones, which became visible with the tempograms. Thirdly, timbre energy also plays a role with the danceability value. Looking at the generated cepstograms, the two tracks with the highest danceability also displayed high energy distributed across all the timbre features. Furthermore, the classifier showed that The Party Playlist was generally louder in the 2nd timbre feature. Lastly, the classifier showed to have been able to predicted which track fits in the right playlist. It has predicted the track right with an 84 % accuracy for The Party Playlist and 78% for The Mellow Playlist.

The classifier seems to do a good job in placing the songs into the right playlist. Spotify thus seems to interpret tracks in a similar way as humans. This could be an explaination for Spotify’s renown track and playlist suggestions.